Update README.md
Browse files
README.md
CHANGED
|
@@ -60,6 +60,17 @@ print(similarities)
|
|
| 60 |
### Matryoshka truncation
|
| 61 |
The model supports shortened embeddings via Matryoshka training. Supported dimensions: 2048, 1536, 1024, 768, 512, 256, 128
|
| 62 |
model = SentenceTransformer("johnandru/ko-vdr-preview", truncate_dim=512)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
|
| 64 |
## Training
|
| 65 |
|
|
@@ -68,10 +79,11 @@ Training used a multi-source Korean/English VDR dataset with hard negatives mine
|
|
| 68 |
|Source| Language| Type
|
| 69 |
| --- | --- | ---: |
|
| 70 |
|NomaDamas/ko-vdr-train-public-v2.0 |Korean |Query–page pairs|
|
| 71 |
-
|whybe-choi/ko-vdr-train-private-v0.1| Korean |Query–page pairs
|
| 72 |
-
|vidore/colpali_train_set| English |Query–page pairs
|
| 73 |
-
|tomaarsen/llamaindex-vdr-en-train-preprocessed| English |Query–page pairs
|
| 74 |
-
|Ko/En text retrieval corpus| Korean + English| Text pairs
|
|
|
|
| 75 |
Hard negatives were mined with Qwen/Qwen3-VL-Embedding-8B using absolute_margin=0.05 and 7 negatives per pair (top sampling).
|
| 76 |
### Loss
|
| 77 |
MatryoshkaLoss(SelfGuideCachedMultipleNegativesRankingLoss) — InfoNCE with cosine similarity (scale=20), cached mini-batches (mini_batch_size=4), and Matryoshka multi-granularity weighting.
|
|
|
|
| 60 |
### Matryoshka truncation
|
| 61 |
The model supports shortened embeddings via Matryoshka training. Supported dimensions: 2048, 1536, 1024, 768, 512, 256, 128
|
| 62 |
model = SentenceTransformer("johnandru/ko-vdr-preview", truncate_dim=512)
|
| 63 |
+
## Model Details
|
| 64 |
+
|Property| Value
|
| 65 |
+
| --- | --- :|
|
| 66 |
+
|Base model| Qwen/Qwen3-VL-Embedding-2B|
|
| 67 |
+
|Fine-tuning method| LoRA (r=32, alpha=32, no dropout)|
|
| 68 |
+
|LoRA target modules| q_proj, k_proj, v_proj, up_proj, down_proj, gate_proj|
|
| 69 |
+
|Embedding dimension| 2048 (Matryoshka: 1536 / 1024 / 768 / 512 / 256 / 128)|
|
| 70 |
+
|Precision| bfloat16|
|
| 71 |
+
|Attention| Flash Attention 2|
|
| 72 |
+
|Max image pixels| 1280 × 28 × 28|
|
| 73 |
+
|Framework| sentence-transformers==5.4.1, peft>=0.19.1|
|
| 74 |
|
| 75 |
## Training
|
| 76 |
|
|
|
|
| 79 |
|Source| Language| Type
|
| 80 |
| --- | --- | ---: |
|
| 81 |
|NomaDamas/ko-vdr-train-public-v2.0 |Korean |Query–page pairs|
|
| 82 |
+
|whybe-choi/ko-vdr-train-private-v0.1| Korean |Query–page pairs|
|
| 83 |
+
|vidore/colpali_train_set| English |Query–page pairs|
|
| 84 |
+
|tomaarsen/llamaindex-vdr-en-train-preprocessed| English |Query–page pairs|
|
| 85 |
+
|Ko/En text retrieval corpus| Korean + English| Text pairs|
|
| 86 |
+
|
| 87 |
Hard negatives were mined with Qwen/Qwen3-VL-Embedding-8B using absolute_margin=0.05 and 7 negatives per pair (top sampling).
|
| 88 |
### Loss
|
| 89 |
MatryoshkaLoss(SelfGuideCachedMultipleNegativesRankingLoss) — InfoNCE with cosine similarity (scale=20), cached mini-batches (mini_batch_size=4), and Matryoshka multi-granularity weighting.
|